Hybrid Method Incorporating a Rule-Based Approach and Deep Learning for Prescription Error Prediction

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Pediatric Pharmacology

Background:

• Prescribing errors pose significant risks in healthcare.
• Current Drug Utilization Review (DUR)-based rule systems for error detection can lead to 'warning fatigue' for clinicians.
• Automated detection methods are emerging as a solution to mitigate these risks.

Purpose of the Study:

• To develop and evaluate a novel hybrid artificial intelligence (AI) method for predicting prescription errors.
• To improve the accuracy of prescription error detection beyond existing rule-based systems.
• To reduce alert fatigue experienced by healthcare providers and enhance the quality of medical care.

Main Methods:

• A retrospective analysis of 15,281 patient-level observations from pediatric prescriptions (January 1 - December 31, 2018) using a common data model (CDM) and DUR data.
• Development of an advanced rule-based deep neural network (ARDNN) model, incorporating inspection information and clinician-defined rules for 35 drugs.
• Data included 137,802 normal prescriptions and 1,609 prescription errors.

Main Results:

• The developed ARDNN model achieved a precision of 72.86%, recall of 81.01%, and an F1 score of 76.72%.
• The hybrid method successfully reduced alarm pop-up alert fatigue to below 10%.
• A comprehensive ARDNN-based dashboard was created for real-time monitoring of prescription errors.

Conclusions:

• The novel ARDNN method represents an advancement over traditional rule-based models for prescription error detection.
• Implementing this deep learning-based approach can significantly improve medical efficiency and patient service quality.
• Reducing clinician fatigue through minimized alert overload is a key benefit of this AI-driven system.